A paired re-annotation of 190 Indian court case summaries under 38 Bharatiya Nyaya Sanhita (BNS) 2023 charge categories, labelled from an investigation-stage police perspective rather than a judicial one.

Dataset Description

PoliceDrishti pairs each of the 190 factual summaries from the Original ACI dataset with a fresh set of BNS 2023 charge labels reflecting what a police officer would charge at the investigation stage.

Split	Documents	Positive links
Train	120	549
Test	70	272
Total	190	821

Charge vocabulary: 38 BNS categories (after label normalisation)
Average charges per document: 4.3 (train), 3.9 (test)
Annotation: Single annotator with law background; 30-document double-annotation subset achieves Cohen's κ = 0.78

Key difference from judicial labels

The original dataset assigns charges based on what the court ultimately convicted. PoliceDrishti assigns charges based on what a police officer would lodge at the FIR/investigation stage, using BNS 2023 sections rather than IPC sections. This produces systematically more charges per document and a larger charge vocabulary.

Dataset Structure

Dataset_JSON/
├── P-POV_Train_Set.json   # 120 documents
└── P-POV_Test_Set.json    # 70 documents

Each JSON file is a dict keyed by case ID:

{
  "1954_S_39": [
    {
      "Fact_ID": "1954_S_39",
      "Case Summary": "The case of the prosecution was that ...",
      "P_POV": ["murder", "cheating", "criminal conspiracy"]
    }
  ]
}

Features

Feature	Type	Description
`Factual Case Summaries`	`string`	Factual summary of the court case
`BNS Charges`	`Sequence[string]`	Police-perspective BNS 2023 charge labels

38 BNS Charge Categories

abetment, attempt to murder, causing death by negligence, causing hurt,
causing miscarriage, cheating, criminal breach of trust, criminal conspiracy,
criminal intimidation, criminal trespass, cruelty by husband, culpable homicide,
dacoity, destruction of property, extortion, forgery, hurt, kidnapping,
malicious intention, marriage offence, mischief, murder, offence affecting public safety,
offence against public justice, offence against state, offence by public servant,
offence related to religion, receiving stolen property, rioting, robbery,
sexual offence, suicide, theft, threatening, unlawful assembly, unlawful custody,
wrongful confinement, wrongful restraint

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("Happyman11/PoliceDristi")

print(ds["train"][0])
# {
#   'Factual Case Summaries': 'The case of the prosecution was that ...',
#   'BNS Charges': ['murder', 'cheating', 'criminal conspiracy']
# }

Data Fields

Factual Case Summaries (string): The factual background of the case as summarised in the court judgment.
BNS Charges (List[string]): List of BNS 2023 charge labels assigned from the police investigation-stage perspective.

Annotation Process

Labels were assigned by a single expert annotator with legal background who:

Read the factual case summary (not the judgment outcome)
Identified which BNS 2023 sections a police officer would invoke at the FIR/investigation stage
Assigned one or more of 38 BNS charge categories

A random subset of 30 documents was double-annotated independently to measure reliability. Cohen's κ = 0.78 on the 30 × 38 = 1,140 binary labelling decisions.

Intended Use

Automatic Charge Identification (ACI) from the police perspective
Cross-perspective transfer: training on judicial labels and evaluating on police-perspective labels (and vice versa)
Legal NLP research under Indian BNS 2023
Studying label shift between judicial and investigation-stage perspectives

Limitations

Labels reflect a single annotator's interpretation of the police perspective; they are expert-produced simulations, not actual FIR records or police officer annotations.
The 38-category vocabulary uses BNS 2023 sections; 3 categories have zero occurrences in the current 190-document corpus.
Case summaries are derived from publicly available Supreme Court of India judgments.

Citation

Note: This dataset accompanies a paper submitted to the EACL 2027 Student Research Workshop. Author information will be updated upon acceptance. Please cite as:

@inproceedings{author2027policedrishti,
  title     = {When Labels Depend on Perspective: A Paired Re-annotation Study
               of Automatic Charge Identification},
  author    = {[To be updated upon acceptance]},
  booktitle = {Proceedings of the EACL 2027 Student Research Workshop},
  year      = {2027},
  note      = {Dataset available at https://huggingface.co/datasets/Happyman11/PoliceDristi}
}

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Downloads last month: -

Total file size:

502 kB